3:08Transcript
Pelosi Trades Are Public PDFs. MapReduce on the Ones OCR Can't Read.
Congressional trade PDFs are public, but many are scans. I run MapReduce on them: OCR and an LLM read each page, and a reducer settles the disagreements.
NEXUSTRADE / VIDEOS
AI trading videos
Shorts I filmed while building NexusTrade.
Every video has a transcript.
184 shorts · every one transcribed
LLM videos
Start watching ↗3:08Transcript
3:22Transcript
Why Smaller Language Models Hallucinate More. I Turned Down $475,000.
Why smaller language models hallucinate more. Senior AI engineer. I turned down $475,000. More weights and more data make a more accurate function.
5:41Transcript
MapReduce for LLMs: How Cheap Models Beat Claude Opus 5 on 100 PDFs
I teach MapReduce for LLMs with a concrete test: is Hello World in these 100 PDFs? Split the task, map it across cheap models on different views of the input, then let one reducer combine their answers.
2:26Transcript
Can AI Improve Qwen by Editing Its Weights?
Direct Qwen3.5-9B weight editing with Codex: the saved checkpoints, random controls, a repetition-loop fix and the actual nine-case grading results.
3:08Transcript
How to Fine-Tune an LLM Judge on a Rented GPU. It Has to Beat DeepSeek.
How I'm fine-tuning an LLM judge with LoRA on one rented H200: five steps, a thirteen-dollar pilot, and a race against base Qwen and DeepSeek.
2:51Transcript
LLM Judge Training Data: My AI Checker Changed 12 of 29 Grades
How I built the answer key for my AI critic: real agent runs, an AI checker that changed 12 of 29 grades, and blind review that left 889 trusted examples.
2:25Transcript
How to Train an AI to Build Trading Strategies: Step One Is a Critic
I'm training an AI agent to build profitable trading strategies. Step one is a critic that grades its work; step two is Qwen 3.5 trained with reinforcement learning.